Professional Documents
Culture Documents
Mechanism Information
Nadya Vasilyeva (Vasilyeva@Berkeley.Edu)
Tania Lombrozo (Lombrozo@Berkeley.Edu)
Department of Psychology, 3210 Tolman Hall
Berkeley, CA 94720 USA
Abstract
We report four experiments demonstrating that judgments of
explanatory goodness are sensitive both to covariation
evidence and to mechanism information. Compared to
judgments of causal strength, explanatory judgments tend to
be more sensitive to mechanism and less sensitive to
covariation. Judgments of understanding tracked covariation
least closely. We discuss implications of our findings for
theories of explanation, understanding and causal attribution.
Keywords: explanation; covariation; mechanism; causal
strength; understanding
Experiment 1a
Experiment 1 presented participants with two factors that
were selected such that they would not suggest an obvious
causal relationship. Participants received evidence about the
covariation between these factors that suggested no
relationship, a weak relationship, a moderate relationship, or
a strong (deterministic) relationship. We also manipulated
whether they received information about a possible
mechanism.
Method
Participants Four-hundred-and-ninety-two participants
were recruited on Amazon Mechanical Turk in exchange for
$1.45. In all experiments, participation was restricted to
users with an IP address within the United States and an
approval rating of at least 95% based on at least 50 previous
tasks. An additional 217 participants were excluded for
failing a comprehension check for covariation tables (18),
failing a memory check (199), or both (27).
Materials, Design, and Procedure Participants first
completed a practice session in which they were introduced
Table 1. Sample covariation matrices from Experiments 1-2.
Conditions correspond to P = .04, .33, .64 and 1.
None (nearly)
Weak
Moderate
Strong
Experiment 1b
Experiment 1a found that explanations were judged better
the stronger the corresponding covariation evidence, and
when a full mechanism was provided. We also found that
explanation judgments were less sensitive to covariation
evidence than were causal judgments, but more sensitive
than understanding judgments. The effect of mechanism did
not differ significantly across judgment types.
In Experiment 1b we tested whether the specification of a
full mechanism was necessary to observe a mechanism
effect, or whether it would suffice to state that some
mechanism connected the two factors. If people suffer from
an illusion of explanatory depth (Rozenblit & Keil, 2002)
and make do with quite skeletal mechanistic understanding
(Keil, 2003), one might anticipate a boost in judgments
from even a mechanism sketch or placeholder, and that this
would be greater for explanation than causal judgments. We
therefore duplicated the structure of Experiment 1a, but
replacing detailed mechanism descriptions with a
mechanism pointer - the statement that the factors in
question are related via some unspecified mechanism.
Method
Participants Four-hundred-and-eighty-two participants
were recruited on Amazon Mechanical Turk in exchange for
$1.45. An additional 198 participants were excluded for
failing a comprehension check for covariation tables (17),
failing a memory check (181), or both (27).
Materials, Design, and Procedure were the same as in
Experiment 1a, with the exception of the mechanism
statement: the full mechanism was replaced with a general
statement that there exists some multi-step pathway
connecting the cause to the effect, omitting all other details.
Mechanism pointer: When designing the survey, the
researchers thought they would be related by a multi-step
pathway connecting being a woman to being hit by a bus
Experiment 2
Although providing detailed mechanisms in Experiment 1a
boosted all ratings, the effect was weaker than we expected,
which could have masked differences across judgments. In
particular, it is possible that by presenting Experiments 1a
and 1b as studies about the way people understand data
tables, taking participants through an extensive practice
session focusing on covariation tables, and manipulating
covariation within subjects (while judgment and mechanism
varied between subjects) we artificially drew attention to the
covariation manipulation at the expense of the mechanism
information. To address these concerns, we conducted
Experiment 2, in which we minimized task features that
drew attention to the covariation tables, hoping that it would
Method
Participants Two-hundred-and-fifty-one participants were
recruited on Amazon Mechanical Turk in exchange for
$1.55. An additional 81 participants were excluded for
failing a memory check.
Materials, Design and Procedure Mechanism information
(none, pointer, full) and covariation strength (none, strong)
were manipulated within subjects, and rotated through items
across participants. The type of judgment (explanation
goodness, causal strength, sense of understanding) was
manipulated between subjects.
The materials and procedure were the same as in
Experiments 1a and 1b, with the following exceptions: the
number of items (cause-effect pairs) was reduced to 6 and
the practice session was shortened, as the comprehension
questions about covariation tables were removed to avoid
pragmatic cues that covariation evidence should be
prioritized over mechanism information during the task. All
questions were presented in the token format.
Method
Participants Ninety-one participants were recruited on
Amazon Mechanical Turk in exchange for $1.00. An
additional 16 participants were excluded for failing a
memory check.
Experiment 3
Focusing on explanation ratings versus causal strength
ratings and on the contrast between no mechanism and a full
mechanism, Experiment 2 produced a double dissociation,
with explanation ratings more sensitive than causal ratings
when it came to mechanisms, and causal judgments more
sensitive than explanation judgments when it came to
covariation. While the differential effect of covariation was
also found in Experiments 1a and 1b, the effect of
mechanism information was not. We therefore sought to
replicate the interactions between mechanism and judgment
in Experiment 2 before drawing strong conclusions. We also
tied the mechanism more closely to each judgment by
embedding the mechanism information in the body of the
explanation and causation statements themselves.
General Discussion
In four experiments we demonstrate that judgments of
explanation goodness are sensitive to both covariation
evidence and mechanism information. Comparing
explanation to other judgments, we observed a consistent
dissociation: explanation judgments were less responsive to
the degree of covariation in the data than were causal
judgments. In contrast, specifying a full mechanism had a
stronger effect on explanations than on causal judgments in
Experiments 2 and 3, which drew less attention to the
covariation tables. Of the three judgment types, sense of
understanding was least responsive to covariation. Overall,
our results indicate that these three types of judgments differ
systematically when it comes to the role of covariation data
and the effects of specifying a full mechanism.
Returning to the issues raised in the introduction, our
findings support some tentative conclusions and raise
additional questions for further study. First, we find that
explanations are judged better when supported by stronger
covariation evidence or by the specification of a
mechanism, and that the benefits of stronger evidence are
not limited to cases in which a mechanism is also specified.
It would be interesting to know whether these two factors
affect explanation ratings for different reasons for
example, covariation might be valuable for purely evidential
reasons, while the specification of a mechanism could be a
genuine virtue in addition to having evidential import.
Second, full mechanism information does appear to have
a larger effect on explanation goodness ratings relative to
causal strength ratings, as might be expected on the view
that explanations are especially geared towards generalization (Lombrozo & Carey, 2006), which full mechanism
information supports. More speculatively, it could also be
that reduced sensitivity to covariation emerges for a similar
reason: a certain degree of resistance to over-fitting the data
from a single sample could help achieve more reliable
generalizations (and indeed, Williams, Lombrozo, &
Rehder, 2013 show that explanation encourages a search for
broad patterns despite inconsistent data).
Third, our findings suggest that explanatory goodness
cannot be reduced, in any straightforward way, to judgments
of causal strength. Similarly, ratings of understanding
diverge from those of either explanation or causation. Our
findings thus call for caution when characterizing one of
these judgments in terms of another, and also raise questions
about the extent to which different kinds of explanatory and
causal judgments could diverge. For instance, evaluating
explanatory goodness could diverge from evaluations of
explanation probability, and evaluations of causal structure
could diverge from those of strength.
In sum, we demonstrate that judgments of causal strength,
explanatory goodness and, to some extent, understanding
respond differently to covariation and full mechanism
information. Explanations surpass causal judgments in their
sensitivity to a full mechanism, and the pattern is reversed
for covariation. Our results present a challenge for proposals
that characterize explanations as identifying causes, and
Acknowledgments
This work was supported by the Varieties of Understanding
Project, funded by the John Templeton Foundation.
References
Ahn, W.-K., & Kalish, C. (2002). The role of mechanism
beliefs in causal reasoning. In F. C. Keil & R. A. Wilson
(Eds.), Explanation and cognition. MIT Press.
Ahn, W. K., Kalish, C. W., Medin, D. L., & Gelman, S. (1995).
The role of covariation versus mechanism information in
causal attribution. Cognition, 54, 299352.
Bateson, M., Nettle, D., & Roberts, G. (2006). Cues of being
watched enhance cooperation in a real-world setting. Biology
Letters, 2, 412414.
Bechtel, W., & Abrahamsen, A. (2005). Explanation: a
mechanist alternative. Studies in History and Philosophy of
Biological and Biomedical Sciences, 36(2), 421441.
Cheng, P. W., & Novick, L. R. (1990). A probabilistic contrast
model of causal induction. Journal of Personality and Social
Psychology, 58, 545-567.
Danks, D. (2005). The supposed competition between theories
of human causal inference. Philosophical Psychology, 18,
259272.
Douglas, H. E. (2009). Reintroducing prediction to explanation.
Philosophy of Science, 76(4), 444-463.
Keil, F.C. (2003). Folkscience: Coarse interpretations of a
complex reality. Trends in Cognitive Science, 7, 368-373.
Koslowski, B. (1996). Theory and Evidence. MIT Press.
Lipton, P. (2004). Inference to the best explanation.
Psychology Press.
Lombrozo, T. (2007). Simplicity and probability in causal
explanation. Cognitive Psychology, 55(3), 232257.
Lombrozo, T. (2012). Explanation and abductive inference.
Oxford Handbook of Thinking and Reasoning, 260276.
Lombrozo, T., & Carey, S. (2006). Functional explanation and
the function of explanation. Cognition, 99(2), 167204.
Machamer, P., Darden, L., & Craver, C. F. (2000). Thinking
About Mechanisms. Philosophy of Science, 67(1), 125.
Newsome, G. L. (2003). The debate between current versions
of covariation and mechanism approaches to causal
inference. Philosophical Psychology, 16(1), 87107.
Park, J., & Sloman, S. (2014). Causal explanation in the face of
contradiction. Memory & Cognition, 42(5), 80620.
Rozenblit, L., & Keil, F. (2002). The misunderstood limits of
folk science: An illusion of explanatory depth. Cognitive
Science, 26, 521562.
Strevens, M. (2008). Depth: an account of scientific
explanation. Harvard University Press.
Williams, J. J., Lombrozo, T., & Rehder, B. (2013). The
hazards of explanation: overgeneralization in the face of
exceptions. Journal of Experimental Psychology: General,
142(4), 1006-1014.
Woodward, J. (2000). Explanation and invariance in the special
sciences. The British Journal for the Philosophy of Science,
51, 197 254.